test(skill-size): the premerge-pipeline waiver bought a ratchet bump instead of an extraction #779

Dropped by the issue-filer's max_new cap on the /premerge run for #768 (SUGGESTION tier, recorded nowhere else).

tests/skill-size.test.sh:75 — the premerge-pipeline waiver now reads 2486 against a 500-line limit. Its own reason text claims the added prose is all read at runtime:

the GROWTH= line shape, the STOP_SELF_REFERENTIAL routing row and the blocked_on_file wave-return contract are all read at runtime, so none can move to references/

That is true of the Constants block, the output-line shape, the decision-table row and the §2a YAML contract. It is not true of the ~90 lines of reader rationale around SKILL.md:1387-1458 — "why that measurement has to exist at all", "Two consecutive rounds, and never one", "Deliberately the coarse proxy", the four-causes list — and references/common-mistakes.md already duplicates three of those four paragraphs almost verbatim.

Cost: the whole 2486-line SKILL.md is loaded eagerly on every /premerge invocation, and the waiver column that exists to track pending extraction was used to raise the ratchet instead.

The constraint that makes this non-trivial: only SKILL.md is loaded eagerly, so text the pipeline acts on at runtime cannot move behind a reference file — extracting it there deletes a rule rather than shrinking the debt. The split has to be rationale-only.

Note this waiver is also currently pinned at exactly the file's line count with zero headroom, which has already forced two rounds of prose-tightening to land unrelated fixes.

Related: PR #768